Surrogate Modelling of Open-Die Forging

Sammanfattning

The mechanical quality of an open-die forged component is governed by its effective plastic strain distribution, which drives void closure, grain refinement, and recrystallisation throughout the workpiece. This distribution is predicted accurately by finite element simulation, but at a computational cost that makes it impractical for the iterative design of forging pass schedules, where many candidate stroke sequences must be evaluated. This thesis investigates whether a fully data-driven surrogate model can reproduce the three-dimensional effective plastic strain field of a single forging stroke at a fraction of that cost. A multilayer perceptron was trained to predict the strain increment at each node from its coordinates, its current strain, and three process parameters, using 400 finite element simulations sampled across the parameter space. Evaluated on 40 held-out simulations, the surrogate reproduced the strain field with a median normalised error (NRMSE) of 0.80 %, a global MAE of 0.0054, and an R2 of 0.9941, generalising consistently across the sampled parameters, with the largest errors confined to the contact surface. Inference took around 130 milliseconds per simulation against a few hours for the finite element model, a reduction of roughly five orders of magnitude.